Edge Swaps Phi-4-mini for Aion, but Its Prompt API Trails Chrome by Six Months

Edge now runs Aion Instruct instead of Phi-4-mini and opens a Prompt API origin trial, six months after Chrome shipped the same API on Gemini Nano. What changes for developers.

Christopher Ort
By i10X News Desk · Reviewed by Christopher Ort · 5 min read
Share
Hands on a thin laptop on a train table with the countryside passing outside the window.
Image: AI-generated, i10X
On this page
  1. Three changes in one release
  2. How Aion compares with Phi-4-mini
  3. Same API, different model: Edge against Chrome
  4. A Google proposal that ships in Edge first
  5. What to do now
  6. i10x Perspective
  7. Still open

Microsoft has made Aion Instruct the built-in small language model of its Edge browser, replacing Phi-4-mini, and has released the weights under the MIT license, the Edge team announced on 19 November. For web developers the bigger step is that the Prompt API, which lets a site talk to that local model, moves from the Canary and Dev channels into an origin trial on live sites. That still leaves Edge about six months behind Chrome, where the same API has been stable since May and runs on Google’s Gemini Nano.

Three changes in one release

  • Aion Instruct is generally available in Edge and on Hugging Face. Microsoft says it is coming to Windows too, and the model card adds that an optimized version paired with a Microsoft content moderation model will become the Windows system model by the end of January 2027.
  • Prompt API origin trial: sites register for a token and can call the browser’s model directly, with the option to fall back to a cloud provider when a device cannot run it.
  • Decisions API developer preview: in Canary and Dev only, sites can ask yes or no, multiple choice and scoring questions about a piece of text and get back probabilities plus a confidence value, computed in one pass without generating any text.

Edge downloads Aion only when a site or feature first needs it. Turning off on-device AI in the settings deletes the model and stops further downloads, and companies can block the download with two enterprise policies, according to Microsoft’s developer documentation.

How Aion compares with Phi-4-mini

Microsoft describes Aion as smaller and faster than Phi with comparable quality for common web tasks. Putting the two model cards side by side shows what that trade looks like:

  • Size: about 2.29 billion parameters against 3.8 billion for Phi-4-mini, roughly 40% fewer.
  • Context: 32,768 tokens against 128,000, so about a quarter of the text fits into one request.
  • Knowledge benchmarks: Aion’s card reports 55.15 on MMLU and 74.91 on GSM8K. Phi-4-mini’s card lists 67.3 and 88.6.
  • Languages: Aion’s card names eight languages including German, French and Japanese, but multilingual text makes up only about 6% of its training data. On multilingual MMLU it scores 45.51, Phi-4-mini 49.3.

Those benchmark figures come from different test setups (the Phi numbers use 5-shot and 8-shot prompting), so they are not a clean head-to-head. The direction still fits what Microsoft writes itself: Aion “was not trained to excel at world knowledge”. It is built for rewriting, summarizing, classifying and structured output on the page in front of the user, not for answering factual questions.

Same API, different model: Edge against Chrome

Both browsers expose the same LanguageModel interface, so one piece of JavaScript can run in either. What answers differs. Chrome 148, released on 5 May 2026, shipped the Prompt API as a stable feature on Gemini Nano, with text, image and audio input, according to Google’s Prompt API documentation and release notes. Edge’s trial runs on Aion, a text model. Comparing both vendors’ documentation:

  • Status: stable in Chrome since May; an origin trial in Edge since 19 November, limited in time.
  • Hardware: Chrome asks for at least 22 GB of free disk space plus a GPU with more than 4 GB of video memory, or 16 GB of RAM and four CPU cores. Edge asks for 20 GB of free space and 3 GB of video memory, and checks a device performance class of Medium or higher.
  • Platforms: both cover Windows 10 and 11, macOS and Linux, Edge on Linux only from version 159 with Vulkan drivers. Chrome also supports Chromebook Plus devices. Neither runs the Prompt API on phones.
  • Labeling: from Edge 158, text generated through the Prompt API carries an invisible watermark that marks it as AI output, except for responses forced into a JSON schema or pattern. Developers who parse or store the output need to account for it.

For developers this means testing twice. A prompt tuned on Gemini Nano will not behave the same way on Aion, and Microsoft itself names XPIA, its metric for indirect prompt injection, among the safety scores it still has to improve. A model that reads arbitrary web pages is exactly where injected instructions show up, so model output should be treated as untrusted input in both browsers.

A Google proposal that ships in Edge first

The Decisions API Edge is testing is not a Microsoft design. Its explainer on GitHub was written by Google’s Chrome Built-in AI team and says it “has not been approved to ship in Chrome”. Edge is therefore the first browser to put the idea in developers’ hands, running on Microsoft’s own model instead of Google’s. Google launched a server version of the same single-pass approach with the Gemini Decisions API on 12 November.

What to do now

  • Web teams: build on feature detection and a cloud fallback. Check for LanguageModel, then for model availability, and route everything else to a hosted model.
  • Use the trial for real data: the Edge token lets you measure how many of your visitors actually get the local model, which no lab test can tell you.
  • Keep tasks small: with 32K tokens and weak world knowledge, Aion suits classification, rewriting and form extraction, not long documents or fact lookups.
  • Enterprise IT: decide on the download policy before the model reaches managed Windows PCs. Edge only fetches the model when at least 20 GB are free and deletes it again when free space drops below 5 GB.

i10x Perspective

The quiet winner here is the web standard, not either model. Chrome and Edge now expose the same interface, Edge’s trial feedback is meant to feed standardization talks, and Microsoft is borrowing Google’s Decisions design. We expect developers to stop writing for one browser’s AI and start writing for a browser API with a cloud fallback, the way they did for WebRTC.

For Microsoft the browser is a test bed for Windows. Aion becomes the Windows system model in early 2027, so every site that tunes prompts for Edge now is also tuning for Copilot+ PCs later. Shipping a 2.3B model with weak knowledge scores is a deliberate bet on everyday tasks over trivia, similar to the hardware-gated approach Google took with Gemini Nano on Android.

Still open

Microsoft has not published an end date for the origin trial or any benchmark against Gemini Nano. Quantized versions of the weights are announced but not yet released, and it is unclear whether Apple’s Safari will adopt the API at all. More Microsoft model news is collected on our Microsoft page.

Christopher Ort

Christopher Ort

Christopher Ort is the editor of i10X News at i10x.ai. He has worked hands-on with AI tools, models and agent workflows for years and leads i10X’s coverage of AI launches, funding and research. He is the author of “The Style Penalty”, an SSRN study on how frontier AI models judge AI-written CVs, covered by heise online and t3n.

Timeline

What happened before

  1. Release

    MAI-Code-1.1-Flash

    256K context

    microsoft.ai (source, opens in a new tab)
  2. Partnership

    Microsoft commits multiple billions to Mistral’s European GPU capacity

    With Microsoft · Compute

    reuters.com (source, opens in a new tab)
  3. Release

    MAI-Thinking-1 (7-model MAI family)

    Flagship reasoning model plus MAI-Code-1-Flash, MAI-Image-2.5, MAI-Transcribe-1.5 and MAI-Voice-2.

    microsoft.ai (source, opens in a new tab)
  4. Partnership

    Microsoft and OpenAI amend deal, ending exclusive licence

    With OpenAI · Cloud

    openai.com (source, opens in a new tab)
  5. Partnership

    Microsoft and NVIDIA to invest up to $15B in Anthropic

    With NVIDIA, Anthropic · Compute · $15B

    Microsoft (up to $5B) and NVIDIA (up to $10B) to invest in Anthropic; Anthropic commits to $30B of Azure compute.

    anthropic.com (source, opens in a new tab)
Full Microsoft timeline ›

Editor's picks

  1. Georgia Tech’s Academic AI Playbook: 20 Lessons From 67 Real Courses, Free to Reuse
  2. Arcfront’s $1B Run Rate: What the AI Data Startup Confirms and What It Doesn’t
  3. GPT-6.2 Astra Tops Blueprint-Bench 2, 70% Above May Leader GPT-5.5
  4. Deutsche Telekom Puts 100 Million Euros Into Mistral for Sovereign AI in Germany
  5. Inkling 2 Mini: Thinking Machines’ Looped Open Model Tops Qwen on Coding Tests